<<<<<<< HEAD

Load libraries

library(lubridate)
## 
## Attaching package: 'lubridate'
## The following object is masked from 'package:base':
## 
##     date
library(ggplot2)
library(dplyr)
## 
## Attaching package: 'dplyr'
## The following objects are masked from 'package:lubridate':
## 
##     intersect, setdiff, union
## The following objects are masked from 'package:stats':
## 
##     filter, lag
## The following objects are masked from 'package:base':
## 
##     intersect, setdiff, setequal, union
library(tidytext)
library(stm)
## stm v1.3.5 successfully loaded. See ?stm for help. 
##  Papers, resources, and other materials at structuraltopicmodel.com
library(stminsights)
======= >>>>>>> 206c52833a0e7671a4f73f0cb421be8270593253

Load and process the text data.

#load all articles with keyword "democracy"
all_news <- readRDS("data/democracy_news.RDS")

# save legit stopwords, and add some
stopwords <- tibble(
  tokens = stopwords::stopwords("ko", source = "stopwords-iso"),
  lexicon = "stopwords-iso") 
stopwords <- stopwords %>% 
  add_row(tokens = c("들이", "하기", "그것", "때문"), 
          lexicon = "stopwords-iso")

# make tidytext format of the data and remove stopwords 
tidy_news <- all_news %>%
  dplyr::select(tokens, Newspaper, Date, Body, Prezparty) %>% 
  dplyr::mutate(article =  row_number()) %>%
  tidyr::unnest(tokens) %>% 
  dplyr::anti_join(stopwords) 
## Joining, by = "tokens"
  # dplyr::filter(tokens != "민주주의") 

# make term-document matrix from the tidytext format (used later for topic models)
news_sparse <- tidy_news %>%
  dplyr::count(article, tokens) %>%
  cast_sparse(article, tokens, n)

Load the Korean sentiment lexicon.

## Parsed with column specification:
## cols(
##   term = col_character()
## )
## Parsed with column specification:
## cols(
##   term = col_character()
## )

Count the sentiment words.

tidy_news <- tidy_news %>% 
  rename(word = tokens)

tidy_news %>%
  inner_join(senti) %>%
  count(word, sort = TRUE) 
## Joining, by = "word"
senti_news <- tidy_news %>%
  inner_join(senti) %>%
  count(Newspaper, Prezparty, sentiment) %>%
  tidyr::spread(sentiment, n, fill = 0) %>%
  mutate(sentiment = positive - negative)
## Joining, by = "word"
# pdf("neg_news.pdf", width=6, height=2)
ggplot(senti_news, aes(Prezparty, sentiment, fill = Newspaper)) +
  geom_col(show.legend = FALSE) +
  labs(x = "governments from 1990 to 2014") +
  facet_wrap(~Newspaper, ncol = 3, scales = "free_x") +
  scale_fill_grey() +
  theme_bw()
<<<<<<< HEAD

=======

>>>>>>> 206c52833a0e7671a4f73f0cb421be8270593253 
# dev.off()

See the top sentiment words.

senti_word_counts <- tidy_news %>%
  inner_join(senti) %>%
  count(word, sentiment, sort = TRUE) %>%
  ungroup()
## Joining, by = "word"
senti_word_counts

Plot the top sentiment words.

senti_word_counts %>%
  group_by(sentiment) %>%
  top_n(10) %>%
  ungroup() %>%
  mutate(word = reorder(word, n)) %>%
  ggplot(aes(word, n, fill = sentiment)) +
  geom_col(show.legend = FALSE) +
  facet_wrap(~sentiment, scales = "free_y") +
  labs(y = "Contribution to sentiment",
       x = NULL) +
  coord_flip() +
  scale_fill_grey() +
  theme_bw()
## Selecting by n
<<<<<<< HEAD

=======

>>>>>>> 206c52833a0e7671a4f73f0cb421be8270593253

Did negativity increase in specific administrations?

senti_news_admin <- tidy_news %>%
  inner_join(senti) %>%
  mutate(admin = case_when(
    Date > "1990-01-01" & Date < "1993-02-24" ~ "1990-1993 Roh TW",
    Date > "1993-02-25" & Date < "1998-02-24" ~ "1993-1998 Kim YS",
    Date > "1998-02-25" & Date < "2003-02-24" ~ "1998-2003 Kim DJ",
    Date > "2003-02-25" & Date < "2008-02-24" ~ "2003-2008 Roh MH",
    Date > "2008-02-25" & Date < "2013-02-24" ~ "2008-2013 Lee MB",
    Date > "2013-02-25" & Date < "2017-03-10" ~ "2013-2014 Park GH"
  )) %>% 
  filter(!is.na(admin)) %>% 
  count(Newspaper, admin, sentiment) %>%
  tidyr::spread(sentiment, n, fill = 0) %>%
  mutate(sentiment = positive - negative)
## Joining, by = "word"
#pdf("neg_admins.pdf", width=6, height=3)
ggplot(senti_news_admin, aes(admin, sentiment, fill = Newspaper)) +
  geom_col(show.legend = FALSE, position = "stack") +
  labs(x = "administrations") +
  facet_wrap(~Newspaper, ncol = 3, scales = "free_y") +
  scale_fill_grey() +
  theme_bw() +
  theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust=1))
<<<<<<< HEAD

=======

>>>>>>> 206c52833a0e7671a4f73f0cb421be8270593253 
#dev.off()

Plot sentiment words as stacked bars by all newspapers.

ggplot(senti_news_admin, aes(admin, sentiment, fill = Newspaper)) +
  geom_col(position = "stack") +
  labs(x = "administrations") +
  # facet_wrap(~Newspaper, ncol = 3, scales = "free_x") +
  scale_fill_grey() +
  theme_bw() +
  theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust=1))
<<<<<<< HEAD

=======

>>>>>>> 206c52833a0e7671a4f73f0cb421be8270593253

Plot the most unique sentiment words (tf-idf).

senti_words <- tidy_news %>%
  inner_join(senti) %>%
  mutate(admin = case_when(
    Date > "1990-01-01" & Date < "1993-02-24" ~ "1990-1993 Roh TW",
    Date > "1993-02-25" & Date < "1998-02-24" ~ "1993-1998 Kim YS",
    Date > "1998-02-25" & Date < "2003-02-24" ~ "1998-2003 Kim DJ",
    Date > "2003-02-25" & Date < "2008-02-24" ~ "2003-2008 Roh MH",
    Date > "2008-02-25" & Date < "2013-02-24" ~ "2008-2013 Lee MB",
    Date > "2013-02-25" & Date < "2017-03-10" ~ "2013-2014 Park GH"
  )) %>%
  filter(!is.na(admin)) %>% 
  filter(sentiment == "negative")  
## Joining, by = "word"
senti_words <- senti_words %>% 
  filter(admin %in% c("2013-2014 Park GH", "1990-1993 Roh TW")) %>%
  filter(Newspaper == "Hankyoreh") %>%
  group_by(admin) %>%
  count(word, sort = TRUE) %>% 
  ungroup() %>%
  bind_tf_idf(word, admin, n) %>% 
  arrange(desc(tf_idf)) %>%
  mutate(word = factor(word, levels = rev(unique(word)))) %>% 
  group_by(admin) %>% 
  top_n(20) %>% 
  ungroup() 
## Selecting by tf_idf
senti_words %>% count(admin)
senti_words %>% 
  ggplot(aes(word, tf_idf, fill = admin)) +
  geom_col(show.legend = FALSE) +
  labs(x = NULL, y = "tf-idf") +
  facet_wrap(admin~., ncol = 3, scales = "free") +
  coord_flip() +
  scale_fill_grey() +
  theme_bw() +
  theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust=1))
<<<<<<< HEAD

=======

>>>>>>> 206c52833a0e7671a4f73f0cb421be8270593253 
# pdf("neg_admins.pdf", width=6, height=3)
# dev.off()

Let’s look at the general words, aside from the sentiment words, to get a better idea of the corpus. Create dataframe with total word counts.

news_words <- tidy_news %>%
    mutate(admin = case_when(
    Date > "1990-01-01" & Date < "1993-02-24" ~ "1990-1993 Roh TW",
    Date > "1993-02-25" & Date < "1998-02-24" ~ "1993-1998 Kim YS",
    Date > "1998-02-25" & Date < "2003-02-24" ~ "1998-2003 Kim DJ",
    Date > "2003-02-25" & Date < "2008-02-24" ~ "2003-2008 Roh MH",
    Date > "2008-02-25" & Date < "2013-02-24" ~ "2008-2013 Lee MB",
    Date > "2013-02-25" & Date < "2017-03-10" ~ "2013-2014 Park GH"
  )) %>% 
  count(Newspaper, admin, word, sort = TRUE) 

total_words <- news_words %>% 
  group_by(Newspaper, admin) %>% 
  summarize(total = sum(n))

news_words <- left_join(news_words, total_words)
## Joining, by = c("Newspaper", "admin")

Show term freqency ranks.

freq_by_rank <- news_words %>% 
  group_by(Newspaper, admin) %>% 
  mutate(rank = row_number(), 
         `term frequency` = n/total)

freq_by_rank

Show term freqency ranks for the most negative administrations.

news_words %>% 
  filter(Newspaper == "Hankyoreh") %>% 
  filter(admin %in% c("2013-2014 Park GH", "1990-1993 Roh TW")) %>% 
  group_by(Newspaper, admin) %>% 
  mutate(rank = row_number(), 
         `term frequency` = n/total)

Which are most unique words for each Newspaper? (tf-idf)

news_words <- news_words %>%
  bind_tf_idf(word, Newspaper, n)
<<<<<<< HEAD 
## Warning in bind_tf_idf.data.frame(., word, Newspaper, n): A value for tf_idf is negative:
## Input should have exactly one row per document-term combination.
======= 
## Warning: A value for tf_idf is negative:
##  Input should have exactly one row per document-term combination.
>>>>>>> 206c52833a0e7671a4f73f0cb421be8270593253 
news_words %>%
  select(-total) %>%
  arrange(desc(tf_idf))

Plot most unique words by each newspaper.

news_words %>%
  arrange(desc(tf_idf)) %>%
  mutate(word = factor(word, levels = rev(unique(word)))) %>% 
  group_by(Newspaper) %>% 
  top_n(15) %>% 
  ungroup() %>%
  ggplot(aes(word, tf_idf, fill = Newspaper)) +
  geom_col(show.legend = FALSE) +
  labs(x = NULL, y = "tf-idf") +
  facet_wrap(Newspaper~., ncol = 3, scales = "free") +
  coord_flip() +
  scale_fill_grey() +
  theme_bw() +
  theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust=1))
## Selecting by tf_idf
<<<<<<< HEAD

=======

>>>>>>> 206c52833a0e7671a4f73f0cb421be8270593253

Plot most unique words by each newspaper for the two most negative administrations.

news_words %>%
  arrange(desc(tf_idf)) %>%
  filter(admin %in% c("2013-2014 Park GH", "1990-1993 Roh TW")) %>% 
  filter(Newspaper == "Hankyoreh") %>% 
  mutate(word = factor(word, levels = rev(unique(word)))) %>% 
  group_by(Newspaper) %>% 
  top_n(20) %>% 
  ungroup() %>%
  ggplot(aes(word, tf_idf, fill = Newspaper)) +
  geom_col(show.legend = FALSE) +
  labs(x = NULL, y = "tf-idf") +
  facet_wrap(admin~., ncol = 3, scales = "free") +
  coord_flip() +
  scale_fill_grey() +
  theme_bw() +
  theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust=1))
## Selecting by tf_idf
<<<<<<< HEAD

=======

>>>>>>> 206c52833a0e7671a4f73f0cb421be8270593253

Now let’s look at dtructural topic models for emotional frames regarding democracy. I analyzed a topic model with 12 topics.

topic_model <- readRDS("data/topic_model.RDS")

# prepare covariates
covariates <- tidy_news %>%
  distinct(article, Newspaper, Date, Body, Prezparty) 
startdate <- as.Date("19900101", "%Y%m%d")
covariates$Date2 <- as.Date(covariates$Date, "%Y%m%d")
covariates$Days <- difftime(covariates$Date2, startdate, units="days") %>% 
  as.numeric()

Show summary output that shows the relative proportions of topics.

plot(topic_model, type = "summary")
<<<<<<< HEAD

Show more words that help you to label the topics.

plot(topic_model, type = "labels")

=======

Show more words that help you to label the topics.

plot(topic_model, type = "labels")

>>>>>>> 206c52833a0e7671a4f73f0cb421be8270593253

Make topic labels (these are mine):

##                          1                          2 
##      "Ideology and ideals" "Civil society, education" 
##                          3                          4 
##    "Public opinion, media"              "North Korea" 
##                          5                          6 
##           "Party politics"      "Law and prosecution" 
##                          7                          8 
##                 "The U.S."               "Government" 
##                          9                         10 
##     "Sentimental humanism"   "Int'l democratizations" 
##                         11                         12 
##            "Civic protest"      "Global human values"
<<<<<<< HEAD

Look into sample articles.

Compare which topics have a higher importance in which words (the choice of each topic pair is set below).

=======

Look into sample articles.

Compare which topics have a higher importance in which words (the choice of each topic pair is set below).

>>>>>>> 206c52833a0e7671a4f73f0cb421be8270593253

Estimate effects with smoothing line

# combine estimates for interaction effects
prep_int <- estimateEffect(1:12 ~ Newspaper * s(Days),
                           topic_model, covariates)

Plot relationship between newspapers and topic.

# pdf("meta_cov.pdf")
par(mar=c(9,9,4,2))
plot(prep_int, covariate = "Newspaper", 
     topics = c(1:12), model = topic_model, 
     method = "difference", cov.value1 = "Chosun", 
     cov.value2 = "Hankyoreh",
     xlab = "More like Hankyoreh ... More like Chosun",
     labeltype = "custom", custom.labels = c("Ideology",
                                             "Civil society, education",
                                             "Public opinion, media",
                                             "North Korea",
                                             "Party politics",
                                             "Law and prosecution",
                                             "The U.S.",
                                             "Government",
                                             "Sentimental humanism",
                                             "Int'l democratizations",
                                             "Civic protest",
                                             "Global human values"))
<<<<<<< HEAD

=======

>>>>>>> 206c52833a0e7671a4f73f0cb421be8270593253 
# dev.off()

# ggsave("plots/meta_cov.png", width = 6, height = 6)

Estimate effects model.

# prepare dataframe with the package stm_insights
effects_int <- get_effects(estimates = prep_int,
                           variable = 'Days',
                           type = 'continuous',
                           moderator = 'Newspaper',
                           modval = "Chosun") %>%
  bind_rows(get_effects(estimates = prep_int,
                variable = 'Days',
                type = 'continuous',
                moderator = 'Newspaper',
                modval = "Hankook"),
            get_effects(estimates = prep_int,
                        variable = 'Days',
                        type = 'continuous',
                        moderator = 'Newspaper',
                        modval = "Hankyoreh"))

labels <- c("Chosun" = "Conservative", "Hankyoreh" = "Liberal")

yearseq <- seq(from = ymd("19900101"),
               to = ymd("20141230"), by = "year")
yearnames <- year(yearseq)

Plot topic prevalences by time.

# plot interaction effects
# pdf("topics_12.pdf", width = 6, height = 5)

effects_int %>% 
 mutate(label = case_when(
   topic == 1 ~ "Ideology",
   topic == 2 ~ "Civil society, education",
   topic == 3 ~ "Public opinion, media",
   topic == 4 ~ "North Korea",
   topic == 5 ~ "Party politics",
   topic == 6 ~ "Law and prosecution", 
   topic == 7 ~ "The U.S.",
   topic == 8 ~ "Government",
   topic == 9 ~ "Sentimental humanism",
   topic == 10 ~ "Int'l democratizations",
   topic == 11 ~ "Civic protest",
   topic == 12 ~ "Global human values"
 )) %>% 
 # filter(topic == c(1,2,4,6,9,11)) %>%
 filter(moderator != "Hankook") %>% 
 mutate(value = as.Date(value, origin = '1990-01-01')) %>%   
 mutate(moderator = as.factor(moderator)) %>%
 ggplot(aes(x = value, y = proportion, color = moderator,
 group = moderator, fill = moderator)) +
 geom_line() +
 geom_ribbon(aes(ymin = lower, ymax = upper), alpha = 0.2)  +
 theme_minimal() + 
 labs(x = 'Year', y = 'Topic Proportion',
      # title = "Frame: Ideology and ideals",
      color = 'Newspaper', group = 'Newspaper', fill = 'Newspaper') +
 # scale_x_date(date_breaks = "3 years", date_labels = "%Y") +
 # theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust=1)) +
 facet_wrap(~label, ncol = 3) +
 scale_fill_grey() +
 scale_colour_grey()
<<<<<<< HEAD

=======

>>>>>>> 206c52833a0e7671a4f73f0cb421be8270593253 
# dev.off()

Plot only selected topic prevalences by time.

# plot interaction effects
# pdf("topics_6.pdf", width = 6, height = 3)

effects_int %>% 
 mutate(label = case_when(
   topic == 1 ~ "Ideology",
   topic == 2 ~ "Civil society, education",
   topic == 3 ~ "Public opinion, media",
   topic == 4 ~ "North Korea",
   topic == 5 ~ "Party politics",
   topic == 6 ~ "Law and prosecution", 
   topic == 7 ~ "The U.S.",
   topic == 8 ~ "Government",
   topic == 9 ~ "Sentimental humanism",
   topic == 10 ~ "Int'l democratizations",
   topic == 11 ~ "Civic protest",
   topic == 12 ~ "Global human values"
 )) %>% 
 filter(topic == c(11,1,6,5,9,2)) %>%
 filter(moderator != "Hankook") %>% 
 mutate(value = as.Date(value, origin = '1990-01-01')) %>%   
 mutate(moderator = as.factor(moderator)) %>%
 ggplot(aes(x = value, y = proportion, color = moderator,
 group = moderator, fill = moderator)) +
 geom_line() +
 geom_ribbon(aes(ymin = lower, ymax = upper), alpha = 0.2)  +
 theme_minimal() + 
 labs(x = 'Year (1990-2014)', y = 'Topic Proportion',
      # title = "Frame: Ideology and ideals",
      color = 'Newspaper', group = 'Newspaper', fill = 'Newspaper') +
 # scale_x_date(date_breaks = "2 years", date_labels = "%Y") +
 # theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust=1)) +
 facet_wrap(~label, ncol = 3) +
 scale_fill_grey() +
 scale_colour_grey()
<<<<<<< HEAD

=======

>>>>>>> 206c52833a0e7671a4f73f0cb421be8270593253 
# dev.off()

Estimate effects without smoothing lines.

prep_point <- estimateEffect(1:12 ~ Newspaper * Days,
                           topic_model, covariates)

Plot relationship between newspapers and topics, but this time I sorted the topics manually by order.

# pdf("meta_cov_linear.pdf", width = 6, height = 5)
par(mar=c(4,7,1,1))
plot(prep_point, covariate = "Newspaper", 
     topics = c(11,1,6,3,9,2,5,12,10,8,4,7), 
     model = topic_model, 
     method = "difference", cov.value1 = "Chosun", 
     cov.value2 = "Hankyoreh",
     xlab = "More like Hankyoreh ... More like Chosun",
     labeltype = "custom", custom.labels = c("Civic protest",
                                             "Ideology",
                                             "Law and prosecution",
                                             "Public opinion, media",
                                             "Sentimental humanism",
                                             "Civil society, education",
                                             "Party politics",
                                             "Global human values",
                                             "Int'l democratizations",
                                             "Government",
                                             "North Korea",
                                             "The U.S."
                                             ))
<<<<<<< HEAD 

# dev.off()

Plot topic prevalences by time, without smoothing lines.

# prepare dataframe with the package stm_insights
effects_point <- get_effects(estimates = prep_point,
                           variable = 'Days',
                           type = 'continuous',
                           moderator = 'Newspaper',
                           modval = "Chosun") %>%
  bind_rows(get_effects(estimates = prep_point,
                        variable = 'Days',
                        type = 'continuous',
                        moderator = 'Newspaper',
                        modval = "Hankyoreh"))

labels <- c("Chosun" = "Conservative", "Hankyoreh" = "Liberal")

yearseq <- seq(from = ymd("19900101"),
               to = ymd("20141230"), by = "year")
yearnames <- year(yearseq)

effects_point %>% 
 mutate(label = case_when(
   topic == 1 ~ "Ideology",
   topic == 2 ~ "Civil society, education",
   topic == 3 ~ "Public opinion, media",
   topic == 4 ~ "North Korea",
   topic == 5 ~ "Party politics",
   topic == 6 ~ "Law and prosecution", 
   topic == 7 ~ "The U.S.",
   topic == 8 ~ "Government",
   topic == 9 ~ "Sentimental humanism",
   topic == 10 ~ "Int'l democratizations",
   topic == 11 ~ "Civic protest",
   topic == 12 ~ "Global human values"
 )) %>% 
 # filter(topic == c(1,2,4,6,9,11)) %>%
 filter(moderator != "Hankook") %>% 
 mutate(value = as.Date(value, origin = '1990-01-01')) %>%   
 mutate(moderator = as.factor(moderator)) %>%
 ggplot(aes(x = value, y = proportion, color = moderator,
 group = moderator, fill = moderator)) +
 geom_line() +
 geom_ribbon(aes(ymin = lower, ymax = upper), alpha = 0.2)  +
 theme_minimal() + 
 labs(x = 'Year', y = 'Topic Proportion',
      # title = "Frame: Ideology and ideals",
      color = 'Newspaper', group = 'Newspaper', fill = 'Newspaper') +
 # scale_x_date(date_breaks = "3 years", date_labels = "%Y") +
 # theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust=1)) +
 facet_wrap(~label, ncol = 3) +
 scale_fill_grey() +
 scale_colour_grey()

======= 

# dev.off()

Plot topic prevalences by time, without smoothing lines.

>>>>>>> 206c52833a0e7671a4f73f0cb421be8270593253